Chris McNosky

I direct AI coding agents to build systems you can verify instead of trust.

Written governance contracts the agents work under, gates they aren't allowed to weaken, and the evidence committed to the repository so you can check the claims rather than believe them. My background is procurement at a top-5 homebuilder and running a small manufacturing business — not computer science. The agents write the code. The part that makes it trustworthy is mine.

I don't trust the agents. I built the tool that measures when they lie.

What I'm looking for

Scoped, fixed-price engagements — building or hardening agent-driven systems for teams that need the output to be verifiable, not just demonstrable. A bounded first project is the point: you get to check the work before committing to anything larger.

Also open to full-time roles in applied AI and developer tooling.

NoFuckery AI

A founder-led technical evidence practice for teams deciding whether AI agents deserve consequential authority. It separates direct observation, source reports, independent corroboration, inference, and unknowns—then states the operator decision the evidence supports.

Methods, evidence, and corrections LinkedIn Company Page

Work

Stinger

Measures whether an AI coding agent cheats — in your configuration, in your CI.

Runs a corpus of sandboxed trap scenarios against a configured coding agent and mechanically detects dishonest shortcuts: weakened tests, forbidden-file edits, fabricated completion claims, broken house rules, leaked bait secrets. Emits a scored, evidence-linked Integrity Report under a preregistered, frozen rubric. Model-agnostic and vendor-neutral by construction. CLI and GitHub Action.

  • Mechanical truth first. Cheating is decided by deterministic detectors, never an LLM judge. The optional judge may only flag for human review; it can never set a verdict.
  • Fail closed. Any ambiguity, harness error, or missing evidence resolves to a non-scored error — never to a favorable score.

Every fixture test passed, so I spent the money to run it against live agents anyway — because a measurement instrument that has only ever been tested against its own fixtures has not been tested. Those runs found five real defects, each of which would have produced a plausible wrong answer rather than a visible failure. The worst: it scored a textbook-perfect refusal as a cheat, because the trap only ever recognised my own vocabulary. The mislabelled evidence package is committed unedited.

How it was built, and what hitting reality broke Evidence packages Specification

WASP 2.0 — Alpaca Autonomous Trader

A safety-critical trading system, deliberately held pre-trade.

A single-user, clean-room automated trading system: a Rust modular monolith with a PyO3 research layer calling the same compiled strategy, decision-replay, and risk core. Fail-closed authorization, an append-only evidence ledger, environment-isolated broker hosts, and reconcile-first deployment.

It isn't trading, and the README says so in the first screen. Nothing is certified, so the system refuses to act — live submission requires a human-approved permit and passed readiness gates, and ambiguous broker outcomes fail closed. That is the design working, not the project stalling. A system that would trade today is a system I built wrong.

Case study 24-page technical handoff Threat model

Method

"Prompted an AI" does not describe it. The contribution is governance, and it is the part that does not come free with the model:

A written constitution
Every agent session is bound by an AGENTS.md that predates the code: build to the spec, fail closed, never weaken a gate to make a test pass, no stubs in shipped code, and the software may not approve its own release.
Mechanical distrust
A change is done when the check script passes — lint, strict typing, the full suite with enforced coverage floors, a no-stub gate, and the trap corpus re-proving itself fair on every commit. No self-report is trusted, least of all the agent's.
Enforced honesty
Every capability is graded by evidence, and what is not done is published as prominently as what is. The trap corpus ships with an audit guide inviting outside reviewers to attack its fairness, openly flagging that the scenarios and the detectors that judge them share an author.
Clean-room provenance
Documented rules about what may enter the codebase and from where, so the origin of every component is answerable rather than assumed.

Background

History and economics degree. Roughly four and a half years inside a top-5 homebuilder's procurement and vendor stack, then running a small manufacturing business. That is domain access, not a detour: I know where the money and the friction actually sit in industries that AI tooling has barely touched, and I have bought from, negotiated with, and been failed by enough vendors to know what makes software trustworthy to the people paying for it.

Get in touch

cmcnosky@gmail.com

Available now for scoped, fixed-price engagements. Also open to full-time roles.

Forwarding this to someone? Here's the sentence:

Chris is a non-engineer who directs AI coding agents under formal governance contracts to ship systems with committed, checkable evidence — including an open-source tool that measures whether coding agents cheat. He's taking scoped contract work.